TFT-multi: simultaneous forecasting of vital sign trajectories in the ICU

TL;DR

This work extends the framework temporal fusion transformer (TFT), a multi-horizon time series prediction tool, and proposes TFT-multi, a global model that can predict multiple vital trajectories simultaneously, and demonstrates the model’s competitive performance and computational efficiency compared to state-of-the-art prediction tools.

Abstract

Trajectory forecasting in healthcare data has been an important area of research in precision care and clinical integration for computational methods. In recent years, generative AI models have demonstrated promising results in capturing short and long range dependencies in time series data. While these models have also been applied in healthcare, most of them only predict one value at a time, which is unrealistic in a clinical setting where multiple measures are taken at once. In this work, we extend the framework temporal fusion transformer (TFT), a multi-horizon time series prediction tool, and propose TFT-multi, an end-to-end framework that can predict multiple vital trajectories simultaneously. We apply TFT-multi to forecast 5 vital signs recorded in the intensive care unit: blood pressure, pulse, SpO2, temperature and respiratory rate. We hypothesize that by jointly predicting these measures, which are often correlated with one another, we can make more accurate predictions, especially in variables with large missingness. We validate our model on the public MIMIC dataset and an independent institutional dataset, and demonstrate that this approach outperforms state-of-the-art univariate prediction tools including the original TFT and Prophet, as well as vector regression modeling for multivariate prediction. Furthermore, we perform a study case analysis by applying our pipeline to forecast blood pressure changes in response to actual and hypothetical pressor administration.

Figures (5)

• Figure 1: Bland-Altman Plots for predictions in the held-out test set across vitals. Each point represents an observed time point. Vertical bars represent the upper and lower prediction bound errors; solid line indicate the mean difference, dashed lines indicate the mean difference $\pm1.96$ times the standard deviation of differences. TFT-multi is well calibrated in mean BP and pulse, but limited in features with more missingness.
• Figure 2: Sample prediction visualization across vital features. Different colors indicate different trajectories: blue indicates observed historical data; orange indicates ground truth; green, red and purple indicate the 10th, 50th and 90th percentile prediction respectively.
• Figure 3: Top features by importance in predicting the 50th percentile trajectory for both static and time series variables. Color gradient reflects importance by weight.
• Figure 4: Violin plot comparing percentage of true trajectory within the estimated bounds across vitals of interest for three methods: prophet, TFT-multi (ours) and TFT. Lines within each plot represent the first, second (median) and third quartile values of the held-out test set in (a) and external validation set in (b).
• Figure 5: Model workflow for simultaneously predicting 3 example variables.